Recent content on Bruce on AI Engineering
A 12-star GitHub repo claims to be “the fastest, cheapest browser automation for Claude Code” — 5x faster and 8x cheaper than the tools Microsoft and Google ship. Google Search Console has been showing me people searching that exact tagline for weeks, and as far as I can tell, nobody has actually installed the thing and checked. So I did: cloned lulzx/ubrowser , fought its install for…
“uv vs conda vs pyenv” is one of the most-searched Python tooling questions of 2026, and it’s a category error — like asking whether you should buy a truck, a toolbox, or a jack. These three tools were never in the same race. pyenv manages interpreter versions . conda manages an entire binary ecosystem that happens to include Python. uv manages packages, environments, and…
My blog gets 8% of its traffic from mobile. For a site about developer tooling that number is low but not absurd, so it sat in a dashboard for months without me looking at it. When I finally looked, it took one script and 79 tokens to find a table rendering 427 pixels wide inside a 390-pixel viewport, plus 35 tap targets smaller than the 44-pixel minimum.
The Short Version I’m Bruce. I write about AI engineering — the kind of stuff you actually need when building with LLMs, not the hype. 10+ years building backend systems (Java, Go, cloud-native). Since 2024, fully focused on AI: coding agents, multi-agent architecture, and making AI tools work in real projects.
The Codex app is OpenAI’s desktop command center for coding agents — and since July 9, 2026, it’s no longer a standalone download but a dedicated Codex mode inside the new ChatGPT desktop app, sitting next to Chat and Work. It runs the same agent as the open-source Codex CLI, bills against your existing ChatGPT plan (every tier, including Free), and its specialty is running several…
Two prompts, one shot — the fastest way to show you what this Seedance 2.0 prompt guide is worth. Here’s the version most people type on day one: A beautiful woman walking down a city street at night, cinematic, stunning, epic camera movement, high quality, 8K What comes back is stock footage. A woman who could be anyone drifts past buildings that could be anywhere. The camera can’t…
The official announcement wants you to walk away with one number — Sol scoring 80 on the Coding Agent Index, 2.8 points clear of Claude Fable 5 — and that number is the most overrated line in the entire GPT-5.6 release. The most underrated line sits well below the fold, phrased like a housekeeping note: the Codex desktop app is gone, folded into ChatGPT. The first claim was produced on…
Three tasks. That’s all it took to burn through 73% of a five-hour usage window — and one of the three never even finished. It happened to Kazike (卡兹克), a Chinese blogger on the $200/month Max 20x plan — the most expensive consumer tier Anthropic sells — during Fable 5’s free window. He said it was the first time he’d ever felt token scarcity; in years of shipping code on Opus…
The strangest thing about “ByteDance just beat Gemini at image generation” — the claim that flooded my feed within 48 hours of Seedream 5.0 Pro shipping on July 8, 2026 — is that ByteDance never made it. The launch one-pager calls the model 「全球第一梯队」: first-tier globally . Not first. The “beats Gemini” framing was bolted on downstream, by people who as far as I can tell…
The company that catches falling rockets with steel chopsticks just paid $60 billion for a code editor. Not a satellite maker, not a defense prime — a VS Code fork. And it’s real: on June 16, 2026, four days after the largest IPO in history, SpaceX announced it would acquire Anysphere, the company behind Cursor, for $60 billion in stock . CNBC confirmed the deal , Cursor CEO Michael Truell…
Verdict first, so you can leave early. If you can’t code and need a working product in front of users this week, pick Lovable — $25/month buys a full-stack app with a real database and auth. If your team ships Next.js on Vercel, pick v0 — it’s the only builder whose output arrives as a reviewable pull request. If you’re a developer who wants a dev environment in a browser tab, or…
The Claude AI free tier in 2026 gives you roughly 15–40 messages per rolling 5-hour window — about 30–100 a day if you spread them out. That’s the number you searched for, so it goes first. The range is wide because Claude meters tokens, not messages: a run of short questions lands you near 40, while a single 40-page PDF with a detailed-summary request can drain the whole window in 5–8…
Make Claude Sonnet 5 your default coding model — that’s the answer. If you searched for the best AI coding model 2026 , here it is in one line: Sonnet 5 beats the flagship Opus 4.8 head-to-head on Terminal-Bench 2.1, the benchmark closest to what a coding agent actually does all day, at 40% of the price — and just $2/$10 per million tokens through August 31. Keep Opus 4.8 as your escalation…
Anthropic cut an agent’s token consumption by 84% on a 100-turn web-search eval — and never touched the model. Same weights, same API. The entire win came from context editing , a change to the control loop the agent runs inside. On the same evaluations, context editing alone was worth a 29% performance lift; pairing it with a memory tool pushed that to 39%. No model upgrade in 2026 has…
CLI + Skills is replacing MCP as the default way to wire agents into tools, and the hardest number in the whole MCP vs Skills debate says the fight is basically over: GitHub’s official MCP server initially spent on the order of 50,000 context tokens just describing its tools — later dieted down to roughly 23,000 — while a SKILL.md file that says “use the gh CLI, here’s how”…
It’s 2:40 a.m. and the loop is on iteration 41. The agent has been failing the same test since iteration 12 — somewhere around iteration 23 it “fixed” the problem by deleting the assertion. Files it rewrote an hour ago are getting rewritten again, in the opposite direction. The terminal keeps scrolling, the meter keeps climbing, and nobody’s watching, because nobody having…
You searched for three answers, so here they are up front. Stanford CS146S (“The Modern Software Developer”) is a ten-lecture course on building software with AI coding agents — prompting, agent anatomy and MCP, context engineering, security, code review, and shipping — first taught by Mihail Eric in Fall 2025. It’s worth self-studying in 2026, with one caveat this whole page…
The first time I pointed OpenSpec and Superpowers at the same feature, I ended up with two spec documents for one checkbox. Superpowers’ brainstorming skill fired on its own and wrote a design doc into docs/superpowers/specs/ ; I ran /opsx:propose anyway and got a second spec in openspec/changes/<id>/proposal.md . Within the hour the two files disagreed with each other, and both looked…
Use Windows Terminal with PowerShell 7 as your default profile and WSL2 for your Linux work — for roughly 90% of developers, that’s the best Windows terminal setup in 2026, and it’s probably already installed on your machine. The one upgrade rule worth memorizing: move to WezTerm when you split your week across Windows, macOS, and Linux and want a single config file that follows you…
Most of the effort going into coding agents right now is aimed at the wrong layer. Engineers are still A/B-testing prompt wording, growing their instruction files line by line, and packing the window with every file the agent might conceivably need — all on the instinct that a better-briefed agent is a better agent. That instinct is the closest thing context engineering has to a founding myth:…
$965 billion — and the punchline is that it might be cheap. Snowflake went public in 2020 at roughly 175 times revenue. Datadog crossed 50x at its IPO. Anthropic’s confidential S-1, filed June 2, 2026, asks for about 20x against a $47 billion run-rate. By the one yardstick Wall Street actually uses to price software companies, the scariest valuation in tech is a more modest ask than a…
Apple WWDC 2026 will go down as the most expensive admission of defeat in the company’s history — a two-hour keynote whose actual headline is that Apple quit trying to win AI. The proof fits in one sentence: the new Siri runs on Google’s Gemini models , Apple Foundation Models v2 is being co-developed with Google, and complex queries route to AFM Cloud Pro — hosted on Google Cloud ,…
Earlier in this Harness Engineering series: Part 1—what it is and a 60-day pipeline retrospective , Part 2—CLAUDE.md best practices , Part 3—sub-agent architecture , Part 4—six layers built in reverse . This piece tackles the question those posts dodged: is the term itself hype, and how long does the window stay open? A new term went from a casual coinage to a canonical engineering category in…
Seven weeks ago, Hermes Agent did not exist on GitHub. Today, April 24, 2026, it has 113,000 stars. Version 0.10.0 shipped 8 days ago with 118 skills, three-tier memory, Tool Gateway, and six message gateways. Two versions ago it was 27,000 stars. The growth curve from 0 to 113K in 7 weeks makes Hermes the fastest-growing open-source agent framework of 2026, by a margin of 3-4x over…
On April 4, Anthropic quietly blocked Claude Code subscriptions from calling OpenClaw. No announcement, no blog post. The community found out when the tools stopped working. The HN thread climbed past a thousand points in 48 hours. On April 16 — 12 days later — Anthropic launched Managed Agents in public beta: an official agent runtime with sandboxing, built-in tools, and a new beta header…
Someone sent me wshobson/agents last week and asked the question I keep hearing about hot AI repositories: “OK, but what does it actually do?” 33.9K stars, 3.7K forks, 184 specialized agents, 150 skills, 98 commands, 78 plugins. The README throws numbers at you. Trending lists put it at the top. Dev.to tutorials tell you to install it. None of them tell you what is actually inside,…
The 6 layers of Harness Engineering are not equal. That sentence breaks the standard advice. Every talk, every framework diagram, every “what is Harness Engineering” explainer puts the 6 layers in a tidy stack and tells you to build them in order: Context → Tools → Execution → Memory → Evaluation → Recovery. Sensible, easy to teach, completely wrong if you actually want a stable agent…
Let me open with the contrarian line. Fish shell 4.6, released March 28, 2026, is the most polished interactive shell I have ever used. It is also the wrong choice for your default shell in the AI agent era. Those two claims are not in conflict. They are the whole point of this review. I have been running fish as a front-end for the last nine months, since the 4.0 Rust rewrite landed in February…
The cheapest AI browser automation in 2026 uses zero AI calls . That sentence sounds wrong. We pay for AI agents precisely so we don’t have to write scripts. But after running browser automation through Claude Code, Codex and a few other harnesses for the past few months, I’ve ended up with the opposite conclusion: the AI is the most expensive part of the loop, and you should remove it…
Most AI agent frameworks make you implement the boring parts: tool execution loops, file permission handling, command timeout management. You spend more time on plumbing than on the actual agent logic. The Claude Agent SDK takes a different approach. Instead of giving you an LLM API and saying “build your own tools,” it exposes the exact same toolchain that powers Claude Code — Read,…
I’ve spent the last six months running Apple Silicon as my primary local AI workstation — Ollama, MLX, ComfyUI, Draw Things, llama.cpp, all day every day, across an M4 Pro Mac mini (48GB), an M3 Max MacBook Pro (64GB), and a friend’s M3 Max Mac Studio (128GB). The conclusion is not what the Apple keynotes would suggest.
Nous Research just shipped Hermes Agent v0.9.0 “the everywhere release” on April 13, 2026. Two months after the initial launch the repo is sitting at 27,000+ GitHub stars, and the two weeks preceding v0.9.0 saw 209 PRs merged and 81 issues closed. That is an unusual cadence for an open-source agent project, and I think it is worth unpacking why this one matters beyond the release-note…
Most OpenClaw tutorials teach you the philosophy. None of them give you the one thing you actually need when the agent misbehaves at 2 AM: exact parameters, exact file paths, and exact skill loading rules . This is that reference. After shipping three production OpenClaw agents and contributing two skills to ClawHub, I have learned that 80% of “agent is broken” tickets come down to…
Almost every “terminal AI coding tools comparison” makes the same structural error: put four tools in a flat table, compare features row by row, conclude “it depends on your needs.” This framing is misleading. The 2026 terminal AI coding landscape has diverged into distinct species . These tools differ in architecture philosophy, pricing model, and target user. Comparing…
This is Part 3 of the Harness Engineering series . Part 1 framed the thesis (Agent = Model + Harness). Part 2 went deep on CLAUDE.md , the single most important feedforward control. This article goes deep on a structural decision most teams get wrong: when to spawn a sub-agent, how to route it, and what it actually costs.
1. The Problem: Your Mac Mini Is Stranded Behind a Firewall You have a Mac Mini sitting in the office, always on, always connected. You want to reach it from home — SSH into it, use it as a proxy, or route your traffic through the office network. The problem: it’s behind a corporate NAT firewall, has no public IP of its own, and IT isn’t going to punch holes in the firewall for your…
To perform an interactive rebase in raw Git, you run git rebase -i HEAD~5 , which opens a temporary TODO file in your editor where you manually rearrange lines, change pick to squash or fixup , save, close, and pray you did not make a typo. To stage individual lines from a file, you run git add -p , which walks you through hunks one at a time—and if the hunk granularity is wrong, you edit a patch…
You know that moment when Claude Code starts repeating instructions you gave it ten minutes ago? Or when it produces code that contradicts a decision it made earlier in the same session? That is what context window saturation looks like—and without monitoring, you have zero warning before it happens. I have been using Claude Code daily for months, and the single biggest improvement to my workflow…
Python’s biggest pain point has never been the language itself — it is the tooling around it. Every Python developer has a horror story: pip install that silently broke another project’s dependencies, pyenv install that failed after five minutes of compilation because of a missing system library, or a requirements.txt that worked on one machine but not another. The root cause is…
The Three Walls Claude Code, OpenSpec and Superpowers Are Built to Break If you’ve used Claude Code or any AI coding tool seriously, the three walls below will be familiar. The point of this article is to map each wall to the tool that actually knocks it down — and to flag the common mistake of thinking one tool can do all three.
A one-person “marketing department” that adds 2,000 customers per month and generates $45K MRR. Not a VC-funded team of twenty. One founder, one recently hired DevRel, and four AI employees that never sleep. Postiz founder Nevo David published a detailed account of how he used Paperclip — an open-source AI agent orchestration platform — to take over his entire marketing operation. He…
On March 31, 2026, security researcher Chaofan Shou discovered that Claude Code v2.1.88’s npm package contained a 59.8 MB source map file that should never have shipped. Inside those 512,000+ lines of exposed TypeScript sat a directory nobody expected: src/buddy/ — five files describing a complete virtual pet system. What Anthropic had planned as their April Fools surprise was out in the…
Anthropic’s Claude has become one of the most capable AI assistants available, but its pricing structure can be confusing. With five consumer tiers, two Max sub-tiers, a Team plan with mixed seat types, and a separate API — choosing the right plan takes real research. This guide breaks down every Claude pricing option available in April 2026, including exact costs, usage limits, and…
Asking “which AI coding tool is the best” in 2026 is like asking whether a hammer is better than a screwdriver. The question reveals a misunderstanding of the problem. I have used all five major AI coding tools — Claude Code, Cursor, GitHub Copilot, OpenAI’s Codex CLI, and Google’s Gemini CLI — daily for the past eight months across three production codebases. The…
The biggest threat to your OpenClaw multi-agent system is not misconfiguration. It is letting LLMs handle orchestration . The typical setup looks elegant on paper: a Supervisor agent receives all messages, uses sessions_send to delegate to specialist agents, collects results, and delivers a unified response. In practice, every routing decision is a full API call, every specialist reply-back is…
Claude Code has three distinct extension mechanisms: MCP (Model Context Protocol), Skills , and Hooks . They look related on the surface, but they operate at fundamentally different layers of the system. Choosing the wrong one means wasted effort, bloated context windows, or brittle automation that breaks when you need it most. This guide provides a deep technical comparison — how each extension…
On March 19, 2026, Cursor shipped Composer 2 with a triumphant blog post. Faster, smarter, cheaper — the usual superlatives. Three days later, a developer on X noticed something peculiar in Cursor’s API configuration: a model identifier reading kimi-k2p5-rl-0317-s515-fast . That single string unraveled a story about transparency, open-source ethics, and the increasingly global nature of AI…
On March 31, 2026, a missing .npmignore entry shipped 512,000 lines of unobfuscated TypeScript to the public npm registry. Within hours, the entire internal architecture of Anthropic’s Claude Code — the agent harness connecting LLMs to tools, file systems, and task workflows — was laid bare for the world to study. Two days later, Claude Code Open Source launched as a clean-room Python and…
This is Part 2 of the Harness Engineering series. Part 1 framed the core identity: Agent = Model + Harness . Part 3 goes deep on sub-agent architecture. This piece sits between them, on the single highest-ROI file you will ever write for a coding agent — and the one most teams write wrong. I have read a lot of CLAUDE.md files on GitHub over the past three months. Most of them fail the same three…
This is Part 1 of the Harness Engineering series. Part 2 goes deep on CLAUDE.md best practices . Part 3 covers sub-agent architecture and model routing . I have been running an AI coding harness in production for 60 days — the pipeline that writes, illustrates, and distributes posts for this blog. The single most surprising finding from that experiment: the model is the least important part of the…